Back to Ecosystem Pulse
EvaluationExperimental
ClawBench
by reacher-z
Real-world benchmark for browser AI agents across 144 live websites
Python
Updated Jul 26, 2026
Share:
Summary
Measures browser-based AI agents on 153 everyday online tasks across 144 live websites. Runs a 5-layer recording pipeline, evaluates results with DOM-match heuristics and an LLM judge to score real-world task success. This work aligns with the Planning Pattern and Model Context Protocol (MCP) Pattern for standardized agent context handling.
Why It Matters
As agents act in real web environments, synthetic tests miss crucial failure modes like fragile DOM interactions and live-site variability. ClawBench brings real-world grounding to agent evaluation, surfacing reproducible task-level failures and measurable progress. That visibility is essential for building agent track records and informing trust decisions in multi-agent systems. This emphasis benefits from Human-in-the-Loop Pattern.
Ideal For
Researchers and engineers evaluating web-navigation agents or comparing agent versions on real-world online tasks. This evaluation approach also aligns with practical reasoning and control patterns such as the ReAct Pattern (Reason + Act).
How It's Used
- Measure web navigation and form-filling success rates on live sites
- Compare agent implementations or LLM backends using reproducible recordings and LLM judging
- Diagnose DOM-specific failure modes and flaky behavior on real-world websites
Topics
agent-evaluationagentic-aiai-agent-benchmarkai-agentsbenchmarkbrowser-agentbrowser-automationbrowser-usechrome-agentchrome-extension+10 more
Similar Tools
agent-playground
Keywords
agent-evaluationreal-world-benchmarkweb-agent-benchmarkagent reliability